Abstract: In this paper, a signed floating-point number operation with multi-bit storage three-dimensional resistive random-access memory (3D RRAM) was presented for complex convolution neutral network (CNN) systems. Comparing with other types of memory, 3D RRAM can not only perform calculations inside the memory, but also possess a higher reading rate and a lower energy consumption, providing a new solution to the bottleneck problem of the Von Neumann architecture. A single RRAM cell can reach a maximum and minimum resistance of 10 GΩ and 10 MΩ, which can be stabilized in multi-level resistance states to store high-bit-width data. The test results show that, the accuracy of the signed floating-point number convolution operation system can reach up to 99.8%, the measured peak reading speed of the 3D RRAM model is 0.529 MHz.